Offshore wind harmonic models delivered under IEC~61400-21-3 are increasingly used in grid studies, yet the validation basis behind those models is often weakened by diverging transmission system operator requirements for impedance scan resolution and emission measurement methodology. This paper examines how those requirements affect whether a delivered model actually preserves the converter behaviour and harmonic emission levels that need to be represented in system studies. The work is based on a 15~MW-class offshore wind turbine programme combining field measurement data, reprocessed emission records, and high-fidelity software-in-the-loop studies using production control software. Validation is performed in two steps. First, impedance scans of the same operating points are compared at different frequency resolutions to determine when narrow-band control features are faithfully captured. Second, identical measurement data are reprocessed with different aggregation windows and sample counts to quantify the sensitivity of P95 Norton current estimates to methodological choices. The results show that coarse impedance scans can miss narrow-band converter features in the 30--150~Hz range that are relevant for stability assessment, whereas an adaptive multi-band scan retains the critical detail while reducing simulation effort by more than 80\%. The results also show that short aggregation windows introduce phase-noise-driven inflation of P95 emission estimates, and that too few measurement windows lead to poor percentile convergence. Together, these effects can make a formally compliant model appear more pessimistic or less representative than the underlying turbine behaviour justifies. The paper is relevant to both TSOs and manufacturers because it shifts the discussion from compliance-by-format to validation-by-evidence. The main conclusion is that harmonic model requirements should specify not only what data to deliver, but also the minimum validated methodology behind that data. A harmonised minimum framework, supported by automated quality checks and, where appropriate, software-in-the-loop cross-validation, can improve confidence in delivered models while reducing unnecessary delivery burden.